Pedestrian Detection Algorithm Based On Multi-feature Fusion

Pedestrian detection,as an important topic in the field of computer vision,has played a major role in various fields such as daily production and life,video surveillance,human-computer interaction,and traffic safety,with the advancement of computer technology and the development of machine vision and other theories.However,in the current complex environment,the influence of pedestrians' own variability and external illumination,occlusion and other factor,the performance of pedestrian detection still cannot meet the actual needs,and further exploration and research are needed.This thesis mainly studies the feature extraction and classifier classification of pedestrian detection.In order to overcome the problem that the feature extraction is not accurate enough,this thesis proposes a pedestrian detection algorithm based on improved feature,improving the existing texture features and color self-similarity features.Then,based on the improved features,a multi-feature fusion pedestrian detection algorithm is proposed.The proposed algorithm extracts comprehensive and accurate pedestrian information and improves the effect of pedestrian detection.First of all,this thesis summarizes the research background and research significance of pedestrian detection technology,as well as the research status.The knowledge related to pedestrian detection algorithm is reviewed,including the basic framework of pedestrian detection,some common feature descriptors,consisting of hand-crafted and learning features.In addition to above,classifiers,data set and performance assessment methods used in pedestrian detection also are included,as well as the difficulties and challenges faced by pedestrian detection tasks,in order to have a comprehensive understanding of pedestrian detection.Then,the current texture descriptor LBP feature and color descriptor color self-similarity feature are improved.The texture feature LBP operator encodes the sign of the difference between the pixels in the neighborhood to obtain the texture feature,but ignores the effect of the amplitude,which is not accurate enough.This thesis improves the feature,besides encoding the sign,the amplitude is encoded in the same way to obtain the comprehensive information.At the same time,the histogram Fourier transform is performed to gain the global rotation invariant feature,further enhancing the effect of texture feature extraction.For the purpose of color self-similarity feature,not only does the ratio of pixel sum be calculated to obtain the similarity,butalso the difference is counted,for feature further expansion and validity.By this way,the repeating feature is removed and the discrimination is improved.The experimental results demonstrate that the improved features has better detection effects.Next,based on the above improved feature,a multi-feature fusion algorithm is proposed to extract the extensive information,involving gradient information,texture information and color information.The fusion method is weighted fusion,the weighting coefficients are improved,calculating from the similarity mean and mean variance between samples,retaining the validity of the original data.Classifier uses the improved intersection kernel support vector machine to further improve the detection performance.The experimental results prove that the multi-feature fusion is obviously better than single feature detection,comparing with the existing multi-feature fusion algorithm,the proposed algorithm also has advantages in advancement of detection effects.Finally,the main research contents of this thesis are summarized,and the shortcomings and the future research directions are pointed out.